Schedule
This fall the Responsible AI series follows a single thread from start to finish: building a research dataset with AI assistance. Across four sessions we move from setting up a durable working environment, to gathering materials responsibly, to turning images into structured data, to assembling it all into a pipeline you can rerun and defend. Sessions run in person in the Commons Library Classroom (D112) and online. Materials from previous semesters are in the Past Sessions archive below.
Getting set up — environments, tools, and avoiding lock-in
How to set up a research computing environment using VS Code, Claude Code, and other agentic coding tools. We'll also work through the habits that outlast any single tool: keeping an engineering mindset that avoids vendor lock-in, staying in your own environment rather than copy-pasting between a chat window and your terminal, and knowing when to follow a tutorial and when to ask an AI assistant for help.
Gathering research materials from the Internet
Practical methods for collecting secondary research materials from the web, along with the judgment calls that determine whether a collection is usable and defensible: what you have permission to gather, how to handle privacy, confidentiality, and data sovereignty, and how to document consent and provenance as you go. We'll also practice recognizing when it's faster and more reliable to consult documentation or ask a direct question than to hand a task to a language model.
From images to structured data
Turning scans and digitized materials into structured data using optical character recognition (OCR) and handwritten text recognition (HTR) tools, then focusing on evaluation: using the structure itself — agreement across a row, cross-checking a name against a tenure date — to catch errors instead of trusting model output at face value. This session frames a shift that runs through the whole series: from writing code to reviewing and evaluating what a model produces.
Crafting a research dataset
Assembling gathered materials and structured data into an actual research dataset or database, building a small end-to-end pipeline you could rerun from scratch, and working iteratively from a simple version to a more complete one rather than trying to get everything right in a single pass. We'll also return to the habits of manual review and code review, and to the harder judgment calls: recognizing when an AI-assisted approach has gotten away from you.
Past Sessions
Browse notebooks, slides, and other materials from previous workshops. Each session includes a read-only preview, a one-click link to run the notebook in Google Colab, and a portable version you can run in any Jupyter environment.
Multimodal AI — Video and temporal understanding
Vision-language models can process video and image series, grounding their responses in time to indicate when particular events or shifts occur in a film. This session explored the use of vision-language models for the analysis and interpretation of moving images.
AI for humanities research?
A collaborative session exploring how large language models and vision-language tools can support humanities scholarship — from working with archival image collections via IIIF to contextual analysis of primary sources.
DiScho Discovery Hours — Translating secondary sources
Three practical approaches to translating scholarly texts: quick paragraph-level translation via Google Translate / DeepL, offline reproducible translation with MarianMT, and context-aware scholarly translation with an LLM. Attendees compared outputs on the same passage.
Multimodal AI — Visual tool calling
Multimodal AI models can include visual tools that enable them to manipulate images or retrieve external information. A zoom tool can focus on a section of a painting; reverse image search retrieves metadata. We also built a custom image restoration tool and covered practical document-to-text workflows.
DiScho Discovery Hours — LLM Steering
An exploration of LLM activation steering — adding abstract concept vectors to a model's hidden state to alter its output. We tinkered with the technique using nnsight and sparse autoencoders, and discussed what it reveals about how models represent concepts internally.
Multimodal AI — Visual reasoning and chain of thought
Recent models can reason about the visual contents of images, "thinking aloud" about meaning and relationships between objects. This capability enables more effective recognition of signs and contextual information within images. We explored how this might further visual analysis, interpretation, and distant viewing.
Location
Commons Library Classroom (D112)
